Lifting a higher order function in Haskell

I'm trying to construct a function of type:

liftSumthing :: ((a -> m b) -> m b) -> (a -> t m b) -> t m b

where t is a monad transformer. Specifically, I'm interested in doing this:

liftSumthingIO :: MonadIO m => ((a -> IO b) -> IO b) -> (a -> m b) -> m b

I fiddled with some Haskell wizardry libs and but to no avail. How do I get it right, or maybe there is a ready solution somewhere which I did not find?

This can't be done generically over all MonadIO instances because of the IO type in a negative position. There are some libraries on hackage that do this for specific instances (monad-control, monad-peel), but there's been some debate over whether they are semantically sound, especially with regards to how they handle exceptions and similar weird IOy things.

Edit: Some people seem interested in the positive/negative position distinction. Actually, there's not much to say (and you've probably already heard it, but by a different name). The terminology comes from the world of subtyping.

The intuition behind subtyping is that "a is a subtype of b (which I'll write a <= b) when an a can be used anywhere a b was expected instead". Deciding subtyping is straightforward in a lot of cases; for products, (a1, a2) <= (b1, b2) whenever a1 <= b1 and a2 <= b2, for example, which is a very straightforward rule. But there are a few tricky cases; for example, when should we decide that a1 -> a2 <= b1 -> b2?

Well, we have a function f :: a1 -> a2 and a context expecting a function of type b1 -> b2. So the context is going to use f's return value as if it were a b2, hence we must require that a2 <= b2. The tricky thing is that the context is going to be supplying f with a b1, even though f is going to use it as if it were an a1. Hence, we must require that b1 <= a1 -- which looks backwards from what you might guess! We say that a2 and b2 are "covariant", or occur in a "positive position", and a1 and b1 are "contravariant", or occur in a "negative position".

(Quick aside: why "positive" and "negative"? It's motivated by multiplication. Consider these two types:

f1 :: ((a1 -> b1) -> c1) -> (d1 -> e1)
f2 :: ((a2 -> b2) -> c2) -> (d2 -> e2)

When should f1's type be a subtype of f2's type? I state these facts (exercise: check this using the rule above):

• We should have e1 <= e2.
• We should have d2 <= d1.
• We should have c2 <= c1.
• We should have b1 <= b2.
• We should have a2 <= a1.

e1 is in a positive position in d1 -> e1, which is in turn in a positive position in the type of f1; moreover, e1 is in a positive position in the type of f1 overall (since it is covariant, per the fact above). Its position in the whole term is the product of its position in each subterm: positive * positive = positive. Similarly, d1 is in a negative position in d1 -> e1, which is in a positive position in the whole type. negative * positive = negative, and the d variables are indeed contravariant. b1 is in a positive position in the type a1 -> b1, which is in a negative position in (a1 -> b1) -> c1, which is in a negative position in the whole type. positive * negative * negative = positive, and it's covariant. You get the idea.)

Now, let's take a look at the MonadIO class:

class Monad m => MonadIO m where
    liftIO :: IO a -> m a

We can view this as an explicit declaration of subtyping: we are giving a way to make IO a be a subtype of m a for some concrete m. Right away we know we can take any value with IO constructors in positive positions and turn them into ms. But that's all: we have no way to turn negative IO constructors into ms -- we need a more interesting class for that.